A vision foundation model for single-cell biology via spatial gene cartography

Traditional single-cell foundation models often treat gene expression as a sequence of tokens, similar to natural language processing. This approach risks losing quantitative expression data and ignores the inherent spatial and functional relationships between genes. The authors ask whether reframing single-cell transcriptomics as a computer vision problem—where cells are rendered as images—can improve representation learning and biological interpretability.

To address this, the authors developed scVision. First, they used Gromov–Wasserstein optimal transport to arrange genes on a fixed 104x104 lattice, ensuring that co-expressed genes are spatially adjacent. This transforms a cell's transcriptome into a continuous-valued image (scImage). They then pretrained a Vision Transformer (ViT) on 72 million human cells using masked image modeling. This frozen encoder was then evaluated on its ability to perform cell-type annotation and gene-program discovery across diverse, held-out datasets without any task-specific fine-tuning.

scVision consistently outperforms existing token-based foundation models and classical baselines in zero-shot cell-type annotation. Notably, it demonstrates high label efficiency, achieving superior accuracy with a single labeled example compared to other models provided with fifty. Because the model preserves spatial relationships, the authors can interpret the attention maps as active gene programs. These programs are biologically meaningful and transferable across tissues; for instance, the model identified a shared myeloid gene program in both kidney macrophages and brain microglia without explicit pathway supervision.

By treating gene expression as an image, scVision bridges the gap between single-cell biology and mature computer vision techniques. This spatial representation not only improves the accuracy of cell classification but also provides a more intuitive way to probe gene-program activity and simulate perturbations. It suggests that the spatial organization of the transcriptome is a critical feature for foundation models to capture, offering a robust alternative to the dominant language-model-inspired architectures.